An efficient electricity theft detection based on deep learning

Electrical theft is a pervasive issue that has detrimental impacts on both utility companies and electrical consumers worldwide. It undermines the economic growth of utility businesses, poses electrical risks, and affects customers' expensive energy bills. Smart grids produce vast quantities of data, including consumer usage data which is crucial for identifying instances of energy theft. Machine learning and deep learning algorithms may use this data to identify instances of energy theft. This research presents a new approach using convolutional neural networks and long-short-term memory to extract abstract characteristics from power consumption data, to improve the accuracy of theft detection for electricity users. We mitigate dataset shortcomings, such as incomplete data and imbalanced class distribution, by using LoRAS data augmentation. The method's efficiency is evaluated by using authentic power usage data obtained from the State Grid Corporation of China. Finally, we demonstrate the competitiveness of our approach when compared to other approaches that have been assessed on the same dataset. During the validation process, we attained a 97% accuracy rate, surpassing the highest accuracy reported in previous studies by 1%. We obtained accuracy values of 98.75%, 95.45%, and 97.7%, along with corresponding recall and F1 scores. The findings indicate that the suggested approach surpasses existing state-of-arts methods.